As consumer smartphone recommerce scales, buyers must quote binding repurchase prices from seller-uploaded multi-view photos and diagnostic screenshots before devices reach the warehouse. Accurate remote valuation requires fusing noisy cosmetic evidence with identity, battery, lock, and channel risk from screenshots and staff forms, since an error in any modality changes both grade and payout. However, existing multimodal resources mainly target general VLM competence, screen reading, or factory defect detection, and largely overlook staff-aligned trade-in fields, dual appearance/staff ground truth, and executable valuation scorecards. This gap stems from two challenges: (1) the lack of large-scale real trade-in data with audited dual annotations and a compact ranking benchmark kept separate from training labels, and (2) the need to jointly reason over unordered multi-view cosmetics and diagnostic screens under operational deduction rules. To address these challenges, we propose PhoneValue-Bench, a dual-GT evaluation suite with tier-gated scorecards, together with PhoneValue-Corpus, a matched weakly labeled training set of over twenty thousand orders and hundreds of thousands of captioned images. We further present ValueAgent, a staged perception-fusion-valuation reference pipeline aligned with production inspection, plus an executable public scorer and submission schema. Experiments show that ValueAgent reaches a measured 84.1/100 on Bench-Full, with 97.0 appearance fill and 54.1 staff-field fill, establishing a practical dual-GT benchmark for multimodal recommerce valuation.
Figures & tables
Dimension
DocVQA
ScreenQA
MMBench
RICO
SSGD
VisA
Kaputt
Ours
Large training / crawl set
△
△
✗
✓
△
△
✓
✓
Compact audited dual-GT bench
✗
✗
✗
✗
✗
✗
✗
✓
Multi-view device cosmetics
✗
✗
✗
✗
△
△
△
✓
Diagnostic screenshot fields
✓
✓
✗
△
✗
✗
✗
✓
Staff / order risk fields
✗
✗
✗
✗
✗
✗
✗
✓
Tier-gated valuation scorecard
✗
✗
✗
✗
✗
✗
✗
✓
Table 1: Closest public resources. Consumer phone recommerce valuation benchmarks are absent; nearby defect or UI resources lack dual-GT scorecards.
Figure 1: PhoneValue construction pipeline: collect trade-in orders, preprocess and quality-gate uploads, create dual vision/staff labels, assemble the training corpus, and package the Check/Full evaluation splits.
Figure 2: ValueAgent pipeline (Stages 1–6). Inputs: multi-view photos, repair-history text, and structured device fields. Stage 1 crop–rotates six views; Stage 2 detects defects and emits a cosmetic grade (mint/minor/major/crack/scrap); Stage 3 tests scene compliance (reject outdoor, pass store); Stages 4–5 use a VLM on About and diagnostic screens to list deduction items; an Expert Knowledge Base (identity, risk, abnormal staff items) supports fusion; Stage 6 computes Base Price from quote category × price table, subtracts deductions, and emits Final Price with grade dashboard A–E.
#
Stage
Operation
In: images I , metadata m→ Out: scorecard y^ , grade g , price p
1
Preprocess
V←CropRotate(I) (six views)
2
Object detection
B,y^grade←Detector(V)
3
Scene compliance
if ¬SceneCompliant(V) : flag / reject
4
VLM (About)
sscreen←VLM_ReadAbout(I)
5
VLM (diagnostic)
sdiag←VLM_ReadDiagnostic(I)
Table 2: ValueAgent inference summary
Figure 3: Dataset and benchmark statistics. (a) Cumulative share of corpus cosmetic grades. (b) Order counts by grade (percentages inside bars). (c) Empirical CDF of ValueAgent Score100 on Bench-Check ( n=500 ) and Bench-Full ( n=335 ).
System
Det
VLM
Fuse
Score 100
Y
S
ValueAgent (full)
✓
✓
✓
84.1
97.0
54.1
w/o Perception-Det
✗
✓
✓
68.8
75.7
52.8
w/o Perception-VLM
✓
✗
✓
79.3
97.0
38.2
w/o Fusion / policy
✓
✓
✗
80.5
94.2
48.6
Det-only (no VLM/Fuse)
✓
✗
✗
74.4
97.0
22.0
Gemini-2.5-Pro
✗
✗
✗
69.9
78.0
51.0
Table 3: PhoneValue-Bench-Full ( n=335 ). Score100=(51⋅Yfill+22⋅Sfill)/73 under Tier-B. Factor columns isolate modules: Det =Stages 1–2 crops+detector, VLM =Stages 4–5 screenshot reading, Fuse =Expert-KB R1–R12 (Appendix C). Rows with mixed ✓/✗are controlled ablations; Direct-VLM is the ✗/✗/✗lower bound.
If sscreen.model is non-empty and disagrees with view-label model, keep screenshot model for S01.
R2
Capacity screenshot priority
sscreen.capacity overrides metadata m when both present and disagree.
R3
Worst-region ceiling
y^grade←maxseverity(side,front,back,edge) in scrap ≻ crack ≻ major ≻ minor ≻ mint order.
R4
Water → S08+S10
If any of {detector water-class, sdiag moisture, staff/repair water lexicon} fires: set S08=true; add water_severe to S10 blockers when severity ≥ major.
R5
Activation lock
If About screen or KB flags activation/carrier lock: S10 adds activation_lock ; may force no-quote.
R6
Channel inactive
If staff channel field blank: mark S09 inactive ( gS09=0 ); do not invent channel from UI language alone.
Appendix
Table 11: Expert-KB / fusion rules used by ValueAgent and documented for external reimplementation.
View
Style
English caption (Corpus)
Computer_screenshot
en_narrative
This About screen capture displays the software-identifiable information for an iPhone 13 Pro with 128GB storage and IMEI 3599***8399, sourced from an unlocked other country/region channel. The YOLO model classifies the image as a Computer_screenshot (0.95 confidence) with no detected defects. The staff assessment notes severe external damage, while the order-level YOLO condition hint indicates severe internal damage.
Right_Frame
en_narrative
This right side frame photo of the iPhone 14 Pro Max (128GB, IMEI 3582***8862) shows the device’s side profile. YOLO weak labels detect three minor side scratches and one side dent; staff assessment of heavy wear aligns with the YOLO condition hint.
Front_A
en_narrative
This front black-screen photo of an iPhone 6S Plus (64GB) shows cracked/delaminated screen cosmetics. YOLO reports Front_A with no active defect counts (face grade Normal), contrasting staff severe external damage.
Bottom_Frame
en_concise
Bottom screw-hole photo of an iPhone 13 Pro Max (256GB, heavy wear). YOLO detects 11 side scratches and 2 side dents.
Left_Frame
en_narrative
Left side frame of an iPhone 13 (128GB) with no YOLO defects; staff “excellent” aligns with the condition hint.
Front_B
en_technical
Front_B of an iPhone 13 Pro Max (256GB, IMEI 3579***1099). YOLO stage-2 identifies 21 screen scratches; staff grade heavy wear.
Appendix
Table 12: Six PhoneValue-Corpus English caption examples. Weak labels only; not leaderboard gold.
Field
Type
Example / notes
order_id
string
PV-FULL-0127
views_present
string[]
Six canonical view names when present
defects
object[]
{view, cls, bbox} boxes for Y04–Y07
detector_grade
enum
{ mint , minor , major , crack , scrap }
model / capacity
string
Normalized for S01/S02
imei_last4
string
Masked last-4; null if unknown (do not invent)
Appendix
Table 13: Minimal submission schema (illustrative order PV-FULL-0127). One row per top-level field.
Field
Example
Notes
y / y_max
58.0 / 73
Active sum / Tier-B max
score_100
79.5
Tier-B fill formula
y_fill / s_fill
94.1 / 45.5
Appearance / staff (%)
item_scores
Y01=7.0, …
Points or inactive
gates
Y09=0, S11=0, …
Tier-B mask
Appendix
Table 14: Public scorer output fields for the same illustrative order (not the dataset mean).
Table 25: ValueAgent Y-fill / S-fill by cosmetic severity on Bench-Full ( n=335 ). Overall means match the main table (97.0 / 54.1).
Variant
Active
Score
Δ
Primary hit
w/o Perception-Det
VLM+Fuse
68.8
−15.3
Y01, Y04–Y07
w/o Perception-VLM
Det+Fuse
79.3
−4.8
S01–S03, S10
w/o Fusion / policy
Det+VLM
80.5
−3.6
S02/S08/S09 conflicts
Det-only
Det only
74.4
−9.7
S-fill → 22.0
Direct-VLM (best)
none
69.9
−14.2
Y < 78, weak crops
w/o Stage-3
full pipe
82.7
−1.4
outdoor Y06 FPs
Appendix
Table 26: Ablation point drops from ValueAgent full (84.1) on Bench-Full. Factors match the main-paper results table and Appendix Table 19 .
Pair
r
Score100 vs. grade ± 1
0.61
Score100 vs. price ± 10%
0.48
Score100 vs. price ± 20%
0.55
Y-fill vs. grade ± 1
0.29
S-fill vs. grade ± 1
0.67
Y-fill vs. S-fill
0.18
Appendix
Table 27: Pearson correlations on Bench-Full (ValueAgent). Price-band rows use nprice=298 .
Item
VA
DV
Δ
Appearance items
Y01
100.0
55.0
+45.0
Y02
99.8
61.0
+38.8
Y03
100.0
64.0
+36.0
Y04–Y07 (pooled)
100.0
48.0
+52.0
Y08
74.7
42.0
+32.7
Appendix
Table 28: Mean per-item agreement (%) on Bench-Full active items (aligned with main-paper item table). VA = ValueAgent; DV = best Direct-VLM (Gemini-2.5-Pro). Y04–Y07 share the pooled defect-F1 rate.
Human values are deep motivational orientations that shape human behaviors. In e-commerce, they reveal the stable drivers behind users' purchase decisions. Compared with short-term interests, consumer values better explain how users evaluate products before purchase. However, consumer values are often implicit in complex and fragmented behavioral trajectories, leaving value measurement from e-commerce behaviors largely underexplored. To this end, we propose the Behavior-to-Value (B2V) task, which aims to identify consumer values from e-commerce behavioral trajectories. Centered on this task, we first construct the E-commerce Consumption Value Taxonomy (ECVT) and introduce B2V-Bench, the first B2V dataset and benchmark, based on anonymized Taobao behavioral logs. B2V-Bench consists of real-world purchase decision episodes, covering 25 types of purchase behaviors, along with corresponding consumer value orientations manifested in each episode. To improve consumer value measurement accuracy, we further present B2V-Verifier, a behavior-to-value measurement model based on Value Verification Tuning, which learns to assess whether behaviors provide sufficient evidence for each value inference. Experiments show that B2V-Verifier outperforms strong LLM baselines, improving multi-label classification by 34%. The dataset and code will be publicly released upon acceptance.
Peixuan Hou, Bin Chen, Li He +4
State Key Laboratory of General Artificial Intelligence, School of Intelligence Science and Technology, Peking University · Alibaba Group
Assessing whether multimodal content aligns with macro-societal values, such as peace, justice, and freedom, has become an increasingly urgent challenge. Existing frameworks are largely confined to safety-oriented taxonomies, text-only psychometric probes, or single-label classification. Therefore, we propose MAVEN, a hierarchical framework for macro-societal value evaluation of multimodal content, grounded in international human-rights instruments and cultural value theory. MAVEN organizes values into 6 primary dimensions and 72 secondary indicators, supporting multi-level quantitative scoring. Building on MAVEN, we construct a human-verified multimodal benchmark and a soft-match metric to evaluate VLMs' assessments across value dimensions. For evaluator optimization, we propose a span-adaptive variant of multi-level preference optimization for evaluator distillation, together with a training-free multi-role consensus strategy at inference time. We evaluate existing open- and closed-source VLMs on our benchmark, revealing shared tendencies and clear differences in macro-societal value judgments. Experiments show that our compact 2B evaluator matches its 8B counterpart in the same family and approaches frontier closed-source VLMs, offering a practical path toward scalable macro-societal value evaluation. Our SA-MDPO implementation and MacroValue-Bench are available at https://github.com/zzzzzzzzjj/MAVEN.
Zijuan Zhao, Zheren Fu, Hou Xia +3
University of Science and Technology of China · State Key Laboratory of Communication Content Cognition, People’s Daily Online
Real applications provide the training setting closest to phone-agent deployment, but are difficult to reset, scale safely, and verify programmatically. Static screenshots and interaction trajectories preserve realistic evidence but cannot generate new experience. We introduce PhoneWorld, a trace-grounded framework that converts such evidence into runnable, resettable, and verifiable Android environments. PhoneWorld induces a usage-weighted interaction skeleton from observed pages, transitions, and state-changing operations; translates it into a behavior-grounded app specification; realizes the specification through an autonomous build--inspect--repair loop; and synthesizes executable tasks with programmatic verifiers. The resulting suite spans 34 consumer-facing apps across 16 domains and supports an audited online benchmark, verified trajectory generation, and online RL through common reset and verification interfaces. Evaluations with diverse general and open-source GUI agents show that PhoneWorld supports reliable end-to-end online interaction and exposes capabilities complementary to AndroidWorld. Controlled SFT experiments further show that PhoneWorld trajectories complement AndroidWorld supervision, transfer across online and offline benchmarks, and become more effective as data volume and app coverage increase. Under a matched RL budget, combining PhoneWorld mock-app rollouts with real-app rollouts improves performance over real-app RL alone on both real-phone tasks and AndroidWorld. Together, these results demonstrate that trace-grounded executable abstraction can bridge realistic mobile behavior and scalable agent learning, turning limited real-app evidence into a growing supply of controllable and verifiable environments for training and evaluation.
Yuxuan Liu, Xin Lai, Junyi Li +21
Tencent Hunyuan · Gaoling School of Artificial Intelligence, Renmin University of China · The Chinese University of Hong Kong, Shenzhen +1